Multi-pose And Multi-expression Face Recognition Research

With the construction and development of the safe city,a powerful video surveillance network system has been established throughout the country.The ability to quickly and accurately determine pedestrian identity information is a key factor of intelligent monitoring network.In Security monitoring system which needs to identify the target pedestrians,however only pedestrian IDs are known,so can not well eliminate their influences of existing methods when facing the uncontrollable factors like multi-expressions,mutli-poses etc.So single-sample,multi-expression,multi-pose,and multi-changing environments have caused face recognition to become a big challenge.Aiming at this problem,the paper designs a face recognition system for single-sample,multi-pose and multi-expressions which utilizes VGGNet network model to extract face depth feature,and then constructs multi-confidence statistical criterion to realize face recognition.The paper utilizes VGGNet model to extract face depth feature,and the feature can well represent multi-expression face images and have good robustness to small pose samples at yaw angle on multi-pose datasets.Because the change of face poses will usually exceed ± 30°in video surveillance,the paper designs a single-sample face recognition system based on multi-confidence statistical criteriato by utilizing the characteristic that the change of same person's face posture is smooth and gradual in the surveillance video.The idea of the algorithm is:firstly the test dataset is divided to different pose-level sets,eg:15°,30°,45°,60° and larger pose dataset;then the training dataset is used to recognize the pose 15° set of the test dataset and the identified samples are used as seed samples;next use these seed samples to recognize the pose 30° set,calculate the confidence,and update the sample dataset with high confident samples;and then use these seed samples to recognize the rest pose set(45°,60°)and update the seed samples.These is the reason that the adjacent level pose sample is high similarity and smaller difference,so can be gradually expanded the seed samples on the test datasets;finally the paper utilize the seed samples to train a classification and recognize the rest larger pose set.At the same time,adding a video tracking verification to correct misrecognition samples and reduces the error rate.Face recognition experimental analysis is performed in the multi-pose dataset and surveillance video dataset,which proves the effectiveness of the algorithm.